2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
5 * Harris-Michael lock-free list-based set
6 * http://www.research.ibm.com/people/m/michael/spaa-2002.pdf
26 static node_t *node_alloc (const void *key_data, uint32_t key_len, uint64_t val) {
27 node_t *item = (node_t *)nbd_malloc(sizeof(node_t));
28 memset(item, 0, sizeof(node_t));
29 // If <key_len> is -1 it indicates <key_data> is an integer and not a pointer
30 item->key = (key_len == (unsigned)-1)
31 ? (void *)TAG_VALUE(key_data)
32 : ns_alloc(key_data, key_len);
37 static void node_free (node_t *item) {
38 if (!IS_TAGGED(item->key)) {
44 static void node_defer_free (node_t *item) {
45 if (!IS_TAGGED(item->key)) {
46 nbd_defer_free(item->key);
51 list_t *ll_alloc (void) {
52 list_t *ll = (list_t *)nbd_malloc(sizeof(list_t));
53 ll->head = node_alloc(" ", 0, 0);
54 ll->head->next = NULL;
58 void ll_free (list_t *ll) {
61 static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, const void *key_data, uint32_t key_len, int help_remove) {
62 node_t *pred = ll->head;
63 node_t *item = pred->next;
64 TRACE("l2", "find_pred: searching for key %p in ll (head is %p)", key_data, pred);
66 while (item != NULL) {
67 node_t *next = item->next;
69 // A tag means an item is logically removed but not physically unlinked yet.
70 while (EXPECT_FALSE(IS_TAGGED(next))) {
72 // Skip over logically removed items.
74 item = (node_t *)STRIP_TAG(item->next);
75 if (EXPECT_FALSE(item == NULL))
77 TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
82 // Unlink logically removed items.
84 TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
85 if ((other = SYNC_CAS(&pred->next, item, STRIP_TAG(next))) == item) {
86 TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
87 item = (node_t *)STRIP_TAG(next);
88 if (EXPECT_FALSE(item == NULL))
91 TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
93 // The thread that completes the unlink should free the memory.
94 node_defer_free(other);
96 TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
97 TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
99 return find_pred(pred_ptr, item_ptr, ll, key_data, key_len, help_remove); // retry
101 if (EXPECT_FALSE(item == NULL))
107 if (EXPECT_FALSE(item == NULL))
110 TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
111 TRACE("l4", "find_pred: key %p val %p", STRIP_TAG(item->key), item->val);
113 // A tagged key is an integer, otherwise it is a pointer to a string
115 if (IS_TAGGED(item->key)) {
116 d = (STRIP_TAG(item->key) - (uint64_t)key_data);
118 int item_key_len = item->key->len;
119 int len = (key_len < item_key_len) ? key_len : item_key_len;
120 d = memcmp(item->key->data, key_data, len);
121 if (d == 0) { d = item_key_len - key_len; }
124 // If we reached the key (or passed where it should be), we found the right predesssor
126 if (pred_ptr != NULL) {
131 TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
134 TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key_data, pred);
142 // <key> is not in <ll>.
143 if (pred_ptr != NULL) {
147 TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
151 // Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
152 uint64_t ll_lookup (list_t *ll, const void *key_data, uint32_t key_len) {
153 TRACE("l1", "ll_lookup: searching for key %p in list %p", key_data, ll);
155 int found = find_pred(NULL, &item, ll, key_data, key_len, FALSE);
157 // If we found an <item> matching the key return its value.
159 uint64_t val = item->val;
160 if (val != DOES_NOT_EXIST) {
161 TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
166 TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
167 return DOES_NOT_EXIST;
170 uint64_t ll_cas (list_t *ll, const void *key_data, uint32_t key_len, uint64_t expectation, uint64_t new_val) {
171 TRACE("l1", "ll_cas: key %p list %p", key_data, ll);
172 TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
173 ASSERT((int64_t)new_val > 0);
176 node_t *pred, *old_item;
177 if (!find_pred(&pred, &old_item, ll, key_data, key_len, TRUE)) {
179 // There was not an item in the list that matches the key.
180 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
181 TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
182 return DOES_NOT_EXIST; // failure
185 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
187 // Create a new item and insert it into the list.
188 TRACE("l2", "ll_cas: attempting to insert item between %p and %p", pred, pred->next);
189 node_t *new_item = node_alloc(key_data, key_len, new_val);
190 node_t *next = new_item->next = old_item;
191 node_t *other = SYNC_CAS(&pred->next, next, new_item);
193 TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0);
194 return DOES_NOT_EXIST; // success
197 // Lost a race. Failed to insert the new item into the list.
198 TRACE("l1", "ll_cas: lost a race. CAS failed. expected pred's link to be %p but found %p", next, other);
203 // Found an item in the list that matches the key.
204 uint64_t old_item_val = old_item->val;
206 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
207 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
208 TRACE("l2", "ll_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
212 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
213 TRACE("l1", "ll_cas: found an item %p in the list that matched the key. the expectation was "
214 "not met, the list was not changed", old_item, old_item_val);
215 return old_item_val; // failure
218 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
219 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
220 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
221 // fails it means another thread either removed the node or updated its value.
222 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
223 if (ret_val == old_item_val) {
224 TRACE("l1", "ll_cas: the CAS succeeded. updated the value of the item", 0, 0);
225 return ret_val; // success
227 TRACE("l2", "ll_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
229 old_item_val = ret_val;
234 uint64_t ll_remove (list_t *ll, const void *key_data, uint32_t key_len) {
235 TRACE("l1", "ll_remove: removing item with key %p from list %p", key_data, ll);
238 int found = find_pred(&pred, &item, ll, key_data, key_len, TRUE);
240 TRACE("l1", "ll_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
241 return DOES_NOT_EXIST;
244 // Mark <item> removed. This must be atomic. If multiple threads try to remove the same item
245 // only one of them should succeed.
247 node_t *old_next = item->next;
250 old_next = SYNC_CAS(&item->next, next, TAG_VALUE(next));
251 if (IS_TAGGED(old_next)) {
252 TRACE("l1", "ll_remove: lost a race -- %p is already marked for removal by another thread", item, 0);
253 return DOES_NOT_EXIST;
255 } while (next != old_next);
256 TRACE("l2", "ll_remove: logically removed item %p", item, 0);
257 ASSERT(IS_TAGGED(item->next));
259 // This has to be an atomic swap in case another thread is updating the item while we are removing it.
260 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
261 TRACE("l2", "ll_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
263 // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
264 // item logically removed for a later call (or some other thread) to physically unlink. By marking the
265 // item earlier, we logically removed it.
266 TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next);
268 if ((other = SYNC_CAS(&pred->next, item, next)) != item) {
269 TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other);
273 // The thread that completes the unlink should free the memory.
274 node_defer_free(item);
275 TRACE("l1", "ll_remove: successfully unlinked item %p from the list", item, 0);
279 void ll_print (list_t *ll) {
281 item = ll->head->next;
283 node_t *next = item->next;
284 if (IS_TAGGED(item)) {
288 if (IS_TAGGED(item->key)) {
289 printf("0x%llx ", STRIP_TAG(item->key));
291 printf("%s ", (char *)item->key->data);
294 item = (node_t *)STRIP_TAG(next);